INTERACT-CMIL: Multi-Task Shared Learning and Inter-Task Consistency for Conjunctival Melanocytic Intraepithelial Lesion Grading

2025-12-27

AI Summary

CMIL grading is clinically challenging due to subtle morphological features and overlapping diagnostic criteria, impeding optimal treatment decisions and melanoma risk stratification. To address this, we propose the first multi-task collaborative model for five-dimensional CMIL pathological assessment—encompassing WHO grade 4/5 classification, horizontal/vertical extension, and cytologic atypia. Our method introduces a novel inter-task consistency constraint loss and a composite partial-supervision mechanism, enabling robust training under incomplete annotations. Leveraging shared feature learning, the multi-head architecture integrates a CNN backbone with a custom Inter-Dependence Loss. Evaluated on 486 expert-annotated patches from a multicenter cohort, the model achieves joint multi-axis inference. Compared to single-task baselines, macro-F1 for WHO grade 4 improves by 55.1%, demonstrating strong agreement with expert consensus. We publicly release the first reproducible digital pathology benchmark for CMIL.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Computer Vision: Multi-modal Vision
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Accurate grading of Conjunctival Melanocytic Intraepithelial Lesions (CMIL) is essential for treatment and melanoma prediction but remains difficult due to subtle morphological cues and interrelated diagnostic criteria. We introduce INTERACT-CMIL, a multi-head deep learning framework that jointly predicts five histopathological axes; WHO4, WHO5, horizontal spread, vertical spread, and cytologic atypia, through Shared Feature Learning with Combinatorial Partial Supervision and an Inter-Dependence Loss enforcing cross-task consistency. Trained and evaluated on a newly curated, multi-center dataset of 486 expert-annotated conjunctival biopsy patches from three university hospitals, INTERACT-CMIL achieves consistent improvements over CNN and foundation-model (FM) baselines, with relative macro F1 gains up to 55.1% (WHO4) and 25.0% (vertical spread). The framework provides coherent, interpretable multi-criteria predictions aligned with expert grading, offering a reproducible computational benchmark for CMIL diagnosis and a step toward standardized digital ocular pathology.

Problem

• Accurate grading of conjunctival melanocytic intraepithelial lesions is difficult due to subtle morphological cues.
• The framework jointly predicts five histopathological axes using shared learning and cross-task consistency.
• It aims to provide coherent, interpretable multi-criteria predictions aligned with expert grading standards.

Innovation

• Multi-head deep learning framework for joint prediction
• Shared Feature Learning with Combinatorial Partial Supervision
• Inter-Dependence Loss enforcing cross-task consistency